Scholarlink

The goal of this project is to detect anomalies in financial market data using an LSTM Autoencoder model created with TensorFlow Keras API. The dataset is made up of daily index values from multiple decades and is divided into two sets: training (80%) and testing (20%). The LSTM Autoencoder is a form of RNN that uses an encoder–decoder architecture to replicate input sequences and identify common patterns in data. Anomalies are discovered by calculating the reconstruction error, which is the difference between the original and rebuilt sequence. A threshold derived from the training error distribution indicates significant deviations from normal data. A threshold derived from the training error distribution identifies large departures from normal data. Data points with reconstruction errors that exceed this threshold are classified as anomalies. The capacity of the model to distinguish between normal and anomalous data items is evaluated using precision, recall, and F1-score. Visualizing training and validation loss, as well as highlighting anomalies, reveals unexpected behavior.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Time Series Anomaly Detection via LSTM Autoencoders: A Predictive Analytics Framework

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • R. Sangeethapriya,
  • R. Harihara Sudhan,
  • M. Umamageshwari,
  • M. Dinesh

摘要

The goal of this project is to detect anomalies in financial market data using an LSTM Autoencoder model created with TensorFlow Keras API. The dataset is made up of daily index values from multiple decades and is divided into two sets: training (80%) and testing (20%). The LSTM Autoencoder is a form of RNN that uses an encoder–decoder architecture to replicate input sequences and identify common patterns in data. Anomalies are discovered by calculating the reconstruction error, which is the difference between the original and rebuilt sequence. A threshold derived from the training error distribution indicates significant deviations from normal data. A threshold derived from the training error distribution identifies large departures from normal data. Data points with reconstruction errors that exceed this threshold are classified as anomalies. The capacity of the model to distinguish between normal and anomalous data items is evaluated using precision, recall, and F1-score. Visualizing training and validation loss, as well as highlighting anomalies, reveals unexpected behavior.